| Target tracking is an important research direction in the field of computer vision,which has a wide range of applications and important values in the military and civilian fields.In recent years,a large number of excellent target tracking algorithms have emerged,and gradually developed two technical routes: correlation filter and siamese network.However,due to the difference of target scenes,the performance of these algorithms is not ideal when they are directly applied to complex scenes.In response to this problem,this paper conducts in-depth research on the problem of long-term target tracking in complex scenes such as occlusion,large-scale rotation,and large-scale change in size,and has achieved the following progress:In order to improve the stability of correlation filter in occlusion scene,a self-closed loop anti-occlusion tracking algorithm is proposed.This paper builds a verification module based on SVM to make judgments on the results of short-time tracker FDSST,and combines the state machine to design a variety of update strategies for the filter.At the same time,SVM is used to build a re-detection module to capture the lost target globally.Experimental results show that the verification module,state machine processing strategy and re-detection module achieve the expected design goals,and the self-closed-loop tracking algorithm has good performance in occlusion scenarios.Aiming at the problem that the traditional feature representation ability is insufficient,and siamese network tracking algorithm ignores the shallow convolutional features,a multi-layer feature fusion tracking algorithm is proposed.In this paper,the Trasformer self-attention mechanism and feature long-distance relationship capture capability are used to enhance the features of target template and search region,and build a correlation response graph with global relationships.Meanwhile,a multi-layer feature fusion structure is proposed to mine the edge information contained in the shallow features of convolutional network.Experimental results show that the multi-layer feature fusion algorithm has good performance in long-term tracking dataset,and has good adaptability to target rotation,scale change,camera shake and field of view switch.Aiming at the video processing task requirements of a photoelectric pod,a dual-light detection tracker with localization hardware is built based on Hi3559 A.Based on deployment platform and application scenario,this paper optimized and extended YOLOv3 and template matching to realize automatic target detection function.The structure and calculation of the anti-occlusion tracking algorithm proposed in this paper are optimized to achieve high frame rate tracking in Hi3559 A platform.Experimental results show that the detector can automatically detect the specified target,the tracker has good adaptability to the target with large scale change,and the tracking speed and accuracy meet the task requirements. |
